This is a point-in-time record of the week of 2026-07-13 to 2026-07-19. It was written from the repository's history for that week and is not revised afterwards, so later pages may supersede what it says.

Week of 2026-07-13 to 2026-07-19

Window: 2026-07-13 00:00:00 to 2026-07-19 23:59:59, local author time Recapped: 2026-08-24 Commits in window: 88 (63 non-merge, 25 merges)

This is a backfill written 36 days after the window closed. It records a week in which the lakehouse stopped being an architectural intention and became a working, repeatedly corrected system: Lakekeeper, Spark, Iceberg, production-derived snapshots and MLflow provenance were joined into one local rehearsal path, then the harder DuckDB-to-Iceberg migration began.

What shipped

Plan 124 — contain browser failures inside the solver

Plan 124 was opened after a browser process consumed enough memory to threaten the host. 2d09363 added container memory, swap and PID limits plus tests and the plan document; 678f31f raised the browser allowance after the first limits proved unstable, and 8354e14 aligned the tests and documentation. None of those three commits names Plan 124 in its own text, so they remain in the unattributed ledger even though their changes are described here.

Plan 112 — prove the open-lakehouse foundation

Plan 112 needed a reproducible experiment substrate before adaptive-refresh policy work could be trusted. The week audited the existing features, stood up an isolated Lakekeeper REST catalog, proved Spark/Iceberg writes and time travel against MinIO, rehearsed a real volatility-feature export, added a one-command local snapshot-to-lakehouse path, and attached MLflow provenance. The eleven commits that name the plan are 67a1d1d, cf92f02, 74ab140, 448ef35, 79050b0, 6c801e0, b9ce29a, c243cfd, 89043b1, 062aaaf and 10bc4c9.

The correction chain matters as much as the first implementation. It fixed Lakekeeper image and health assumptions, bootstrapped the empty catalog, created the CI bucket before warehouse registration, moved Spark from S3A to Iceberg's S3FileIO, made cleanup use the catalog's real UUID path, removed an x86-only JAVA_HOME, and separated the production analytics volume from the CI Compose path. Those fixes are named in the unattributed section rather than silently absorbed into the plan count.

Plan 120 — make production-derived fixtures portable

Plan 120 separates expensive snapshot production from Plan 112's consumption of stable fixtures. Gate E packaged and uploaded fingerprinted archives, and Gate F added an authenticated ops/admin download path with alias and manifest cross-checks. The three attributed commits were 92d610c, c1e7195 and the shared local-rehearsal commit 89043b1. Follow-ups hardened the archive prefix, identity checks, OAuth bypass for the authenticated download, MinIO dependency, cache reuse and error handling.

Plan 125 — begin making Iceberg the analytics contract

Plan 125 exists because Plan 112 proved an Iceberg sidecar but left DuckDB as the real analytics contract. The week kept Lakekeeper behind catalog-neutral configuration, audited 22 models and their readers, proved dbt-spark against Iceberg, ported the ten-model volatility chain to exact parity, verified late-arrival behavior, and measured rather than guessed the first Gate C shape decisions.

The work also withdrew a lakehouse CI job after its cold-start fixture problem survived an attempted patch; the local and VM proofs remained valid, but no longer masqueraded as CI coverage. The thirteen attributed commits were 066be10, e063f94, 76aac3f, 498f7a6, 0bb36c8, 96926e6, 5bae544, 3f75283, 430fd15, b63c4f6, 863a2f2, ff33c99 and 9f34557.

Two future streaming documents also entered the record: Plan 126 framed a small replayable event backbone around the existing staging-event seam, and Plan 127 reserved adaptive scrape control for after that transport and its inputs are stable. Their authoring commits (396911a, ac0f2c9) name no plan number and therefore remain unattributed below.

What moved between states

Four revisions of docs/PLANS.md changed the index during the window:

The state changes match the architectural handoff: Plan 125 replaced Plan 118's broader migration placeholder, while Plan 126 replaced the Kafka-shaped Plan 87 with an event-semantics-first design.

What is still owed

• Plan 112 paused after foundation proof; adaptive-refresh model and policy work still depended on Plan 125 making Iceberg a real analytics contract.
• Plan 120 had shipped archive delivery, but its later completion gates and production verification were not closed in this window.
• Plan 125 Gate C still owed production-scale evidence and the actual reader cutover. The week ended with partition specifications decided but not applied.
• Plans 126 and 127 were future designs only. The transport plan depended on a stable Plan 125 substrate; adaptive scrape control depended on the transport and on settled feedback semantics.
• Plan 124 had limits in code, but production verification and the corrected distinction between host-wide and container-local OOMs came later.

Unattributed commits

Thirty-seven non-merge commits name no numbered plan in their own subject or body. Each is included in the work above or described by its subject here; the missing fact is the durable plan attribution, not the work.

Merges

Twenty-five merges are named and not recapped as separate work.

Deferred to the next recap

241 commits sit after this window's end as of 2026-08-24.

git rev-list --count --all --since='2026-07-19 23:59:59'   # 241

That raw tail includes the July 20–26 recap being written alongside this one and the already-covered weeks ending August 2, 9, 16 and 23. Once both files in this backfill are present, none of the following complete weeks is genuinely uncovered.